Edges to Shoes with cGAN (Pix2Pix)

This notebook contains a TensorFlow 2.2.0 implementation (with some comments and explanations) of conditional GAN (cGAN, Pix2Pix) for transforming edges to shoes images from scratch as shown below:

Training Data:

Models are trained on UT Zappos50K shoe image dataset. In this notebook I used a square version of the data.

Images' edges were retreived with scikit-image package, and then used as source images for generation process.

Overall architecture follows the one described in original pix2pix paper:

• U-net as a generator
• PatchGAN as a discriminator

Please check out this article for more detailed implementation guide.

Model Outputs vs Ground Truth:

More Generated Shoes:

Handwritten Sketches to Shoes:

Examples below are much less realistic probably due to my poor painting skills.